Segmenting Narrative Text into Coherent Scenes

This paper describes a quantitative indicator for segmenting narrative text into coherent scenes. The indicator, called the lexical cohesion pro le (LCP), records lexical cohesiveness of words in a xed-length window moving word by word on the text. The cohesiveness of words, which represents their coherence, is computed by spreading activation on a semantic network. The basic idea of LCP is: (1) if the window is inside a scene, the words in it tend to be cohesive, and (2) if the window is crossing a scene boundary, the words in it tend to be incohesive. Comparison with the scene boundaries marked by a number of subject shows that hills and valleys of the graph of LCP closely correlates with the human judgments. LCP may provide valuable information for text analysis, especially for resolving anaphora and ellipsis.